| Summary: | 碩士 === 元智大學 === 電機工程學系 === 106 === This thesis presented a model called as cerebellar model extreme learning machine, and then it is applied to forecasting problems. Because of importance for decision making, forecasting problems always attract researchers. However, nowadays, with the high development of the social economy and technology, decision making has tended to more complicated, and forecasting problem has tended to high chaotic and nonlinear characteristics. The linear model in traditional forecasting method cannot meet the requirement of reality. Artificial neural network has been successfully applied to predict problems in many fields due to its nonlinear fitting ability and good generalization ability. However, the uncertainty of the model and dataset has not been considered in most of the previous researches which focus on point forecasting. The error of traditional point forecasting results is unavoidable because of the uncertainty. For a forecasting result with more reference value, researchers start to consider the uncertainty of the predict model and dataset, then construct the predict intervals for forecasting result. This forecasting method is called as probabilistic forecasting.Extreme learning machine has been widely used in probabilistic forecasting because of the fast speed and good generalization ability, but the shortness of extreme learning machine in terms of accuracy limits its performance. There is a model called as cerebellar model neural network which has excellent nonlinear fitting ability, but its computation speed based on gradient descent method still cannot meet the requirement of probabilistic forecasting.
Based on the above reasons, the main research contents of this thesis include:
(1) A method, called as cerebellar model extreme learning machine (CELM), is suitable for probabilistic forecasting and with higher accuracy and stability is present. CELM is used with Bootstrapping technique to implement the uncertainty and to construct predict intervals.
(2) In addition, wavelet decomposition is used in data preprocess for a higher accuracy.
(3) Effective performance of the proposed model is validated by testing on two applications coming from financial field and industrial engineering field, respectively, including the stock forecasting which data comes from Taiwan securities exchange, and the electric load forecasting which data comes from NingDe power system.
|
|---|
